10 - Clark (Eng) PDF

The British Society for the Philosophy of Science
Connectionism, Competence, and Explanation

Author(s): Andy Clark
Source: The British Journal for the Philosophy of Science, Vol. 41, No. 2 (Jun., 1990), pp.
195-222
Published by: Oxford University Press on behalf of The British Society for the Philosophy of
Science
Stable URL: http://www.jstor.org/stable/687772 .
Accessed: 05/02/2015 22:08
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Oxford University Press and The British Society for the Philosophy of Science are collaborating with JSTOR to
digitize, preserve and extend access to The British Journal for the Philosophy of Science.
http://www.jstor.org
This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

All use subject to JSTOR Terms and Conditions
Brit. J. Phil. Sci. 41 (1990), 195-222 Printedin GreatBritain
Connectionism, Competence, and

Explanation
ANDY CLARK
ABSTRACT
A competence model describes the abstract structure of a solution to some problem,
or class of problems, facing the would-be intelligent system. Competence models
can be quite detailed, specifying far more than merely the function to be computed.
But for all that, they are pitched at some level of abstraction from the details of any
particular algorithm or processing strategy which may be said to realize the
competence. Indeed, it is the point and virtue of such models to specify some
equivalenceclass of algorithms/processing strategies so that the common properties
highlighted by the chosen class may feature in psychologically interesting
accounts. A question arises concerning the type of relation a theorist might expect
to hold between such a competence model and a psychologically real processing
strategy. Classical work in cognitive science expects the actual processing to
depend on explicit or tacit knowledge of the competence theory. Connectionist
work, for reasons to be explained, represents a departure from this norm. But the
precise way in which a connectionist approach may disturb the satisfying classical
symmetry of competence and processing has yet to be properly specified. A
standard 'Newtonian' connectionist account, due to Paul Smolensky, is discus-
sed and contrasted with a somewhat different 'rogue' account. A standard
connectionist understanding has it that a classical competence theory describes an
idealizedsubset of a network's behaviour. But the network's behaviour is not to be
explained by its embodying explicit or tacit knowledge of the information laid out in
the competence theory. A rogue model, by contrast, posits either two systems, or
two aspects of a single system, such that one system does indeed embody the
knowledge laid out in the competence theory.
1 Scenesetting
2 Levelsof explanationand the idea of an equivalenceclass
3 Theclassical cascade
4 Newtoniancompetence
5 Roguecompetence
6 Themethodologyof connectionistexplanation
7 Conclusions:the cascade,the dam and the dividedstream

196 Andy Clark
I SCENE SETTING
In the old days,we all knew what it meantto describethe mindas a syntactic
engine.A syntacticenginewas a physicalsystemcleverlydesignedso that the
way someofits physicalstatesgaveway to otherphysicalstateswas alwaysin
stepwith the way that goodinferencesproceededin some particulardomain.
Forexample,somestatesmightbeusedto standfora categorysuch as dog,and
the physicalsystem set up so that those states reliablygave way to others
which couldbe interpretedas standingforsub-and super-ordinate categories
(such as 'Fido'and 'Mammal').
We understoodthat such an effect(themirroringof semanticregularitiesin
syntactic systems) was made possible by the system's being geared to
manipulatesymbolsaccordingto rules.Symbolswererecurrentphysicalstates
whichwe couldinterpret(e.g.as standingfordog)andthe systemcouldeither
embodythe rulesexplicitlyor implicitly(see the text).
Connectionistsystems(see Rumelhartand McClelland[1986], Smolensky
[1988], Clark[1989]) appearto offera somewhatdifferentway of ensuring
that a physicalsystem is semanticallywell-behaved.In (highly distributed
Smolensky-style)connectionistmodels, there are often no neat recurrent
physicalstates which code for the real world entities which the system is
dealing with. Insteadof being a syntacticengine in which semanticgood
behaviouris ensured by having the system directlyimplementsymbolic
descriptionsof the objectsand processeswhich its inferencesconcern,the
(Smolensky-style)connectionist opts for a statisticalengine operatingon
computationalobjectswhich do not neatlystandforthe objectsandprocesses
in the domain.(Theseobjectsareoftencalled'subsymbols'.) Nonetheless,in a
centralclass of cases, the systembehavesas if it were a symbolic/syntactic
engine. (For a particularlyclear account of this proposal,see Smolensky
[1987] pp. 137-49.)
In what follows I exploresome implicationsof this novel way of being
semanticallywell-behaved.In particular,I askhow well a standardmodelof
explanationin cognitivescience(Marr's3-levelmodel)describesthe connec-
tionist'sprocedureandtheory,andwhetherafailureto fitsuch a modelimplies
a lack of explanatorypower.I begin, then, with some generalcommentson
explanation.
2 LEVELS OF EXPLANATION AND THE IDEA OF AN EQUIVALENCE

CLASS
Explanation,it seems,is a many-levelledthing.A singlephenomenonmay be

subsumedundera panoplyof increasinglygeneralexplanatoryschemas.On
the swingsand roundaboutsof explanation,we tradethe detaileddescriptive/
explanatory power of lower levels for a satisfying width of application at higher

Connectionism,Competence,and Explanation 197
ones. And at each such level there are virtues and vices; some explanations
may be available only at a certain level; but individual cases thus subsumed
may vary in ways explicable only by descending the ladder of explanatory
generality.
For example, the Darwinian, or neo-Darwinian theory of natural selection is
pitched at a very high level of generality. It pictures some very general
circumstances under which 'blind' selection can yield apparently teleological
(or purposeful)evolutionary change. What is requiredfor this miracle to occur
is differential reproduction according to fitness and some mechanism of
transmission of characteristics to progeny. The virtue of this top-level
explanation, then, lies in its covering an open-ended set of cases in which very
differentactual mechanisms(e.g. of transmission) may be involved. In this way
it defines an equivalenceset of mechanisms, that is, a set of mechanisms which
may be disparate in many ways but which are united by their ability to satisfy
the Darwinian demands.
The natural accompaniment to virtue is, of course, vice; and the vice of the
general Darwinian account is readily apparent. We do not yet know, in any
given case, how the Darwinian demands are satisfied. That is to say, we do not
yet have the foggiest idea of the actual mechanisms of heritability and
transmission in any given case. Moreover, there may well be facts about some
specific class of cases (e.g. recessive characteristics in Mendel's peas) which are
not predicted by the general Darwinian theory, which gives us still further
reason to seek a more specific and detailed account.
Mendelian genetics offersjust such an account. It posits a class of theoretical
entities (genes, as they are now called) controlling each trait, and describesthe
way such entities must combine to explain various observed facts concerning
evolution in successive generations of pea plants. The specification included,
for example, the idea of pairs of genes (genotypes) in which one gene may be
dominant, thus explaining the facts about recessive characteristics. (For an
accessible account of evolutionary theory and Mendelian genetics, see Ridley
[1985].)
We may note in passing that between any two levels (e.g. Darwinism and
Mendelian genetics) there will almost certainly be other, theoretically
significant levels. Thus Mendelian inheritance is in fact an instanceof a more
general mechanism called Weismannist inheritance (see Ridley [198 5], p. 23.)
But Weismannist inheritance is still less general than Darwinian inheritance.
Weismannism carves off a theoretically unified subset of general Darwinian
cases. And Mendelism carves off a theoretically unified subset of Weisman-
nism. At each stage the equivalence class is strategically redefinedto exclude a
number of previous members. We can visualize this as a gradual shrinking of
the size of the equivalence class, although this may not be strictly true, since
each new class has a possible infinity of members and so they are, I suppose,
identical in size!

198 Andy Clark
Mendelian genetics provides an interesting case for one further reason. It
was originally conceived as neatly specifyingthe details of lower level DNA-
based inheritance (i.e. of the hardware realization of an inheritance mecha-
nism). As Dennett puts it, Mendelian genes were seen as specifying:
the language of inheritance,straightforwardlyrealised in hunks of DNA
(Dennett[forthcoming]p. 3).
This corresponds to what we shall be terming the classicist vision of the
relation between a certain level of abstract theorizing in cognitive science
(competence theorizing) and actual processing strategies.
But in fact, according to Dennett:
therearetheoreticallyimportantmismatchesbetweenthelanguageof 'bean-bag
genetics'and the molecularand developmentaldetails-mismatches serious
enough to suggestthat all things considered,theredon't turn out to be genes
(classicallyunderstood)at all (Dennett[forthcoming]
p. 3).
This looks like (and is regarded by Dennett as) an analogue of the
connectionist's view of the fate of the constructs of a classical competence
theory.
Be that as it may, the point for now is simply this, that beneath the level of
Mendelian genetics there is some further level of physical implementation
(with God knows what in between, as remarked earlier), and this completes
our descent down the ladder of explanatory generality. We start at the top level
(level 1) with the general Darwinian theory defining a large and varied
equivalence class of instantiating mechanisms. We descend to a more detailed
specification of a subclass of mechanisms (Mendeliantheory) and thence, one
way or another, to the details of the implementation of those mechanisms in
DNA (Figure 1). The effect is a kind of triangulation upon the actual details of
earth-animal inheritance from much broader explanatory principles govern-
ing whole sets of possible worlds.
Explanation in cognitive science, as conceptualizedby, among others, Marr,
Chomsky and Newell and Simon, has a similar multi-layered structure. For a
given class, or class of tasks (e.g. vision, parsing etc.) there will be a top level
story which comprises 'an abstract formulation of whatis being computed and
why', a lower level which specifies a particular algorithm for carrying out the
computation, and (still lower) an account of how that algorithm is to be
realized by physical hardware. To illustrate this, Marr gives the example of
Fourieranalysis. At the top level we have the general idea of a Fourieranalysis.
This can be realized by several different algorithms. And each algorithm in
turn can be implemented in many different kinds of hardware organization
(see Marr [1977], p. 129).
There is an important gap between the 'official' account of the top level
(level 1 as Marrcalls it) and the actual practice of giving 'level 1' theories. For
although the official line is that a level 1 account specifies only the what and

Darwinianequivalence class
Mendelian genetics
DNA
FIGURE 1 The swings and roundaboutsof explanatorygenerality.WhatDNA-based

storiesgain in detailedpowerthey lose in cross-worldscope.
the why of a computation, this specification can be progressively refined so as

to define a more informative (i.e. more restrictive) equivalence-class. This more
refined version of level 1 theorizing (which yet falls short of a full algorithmic
account) has been persuasively defended by Christopher Peacocke under the
title of 'Level 1-5' (see Peacocke [1986]).
The contrast Peacocke highlights is between an equivalence class generated
by defining a function in extension(i.e. by its results-the what, in Marr'sterms)
and a more restrictive (and informative) equivalence class generated by
specifying the body of informationupon which an algorithm draws. Thus, to
adapt one of Peacocke's own examples, suppose the goal of a computation is to
compute depth D and physical size P from retinal size R. And suppose, in
addition, that this computation is to occur inside a restricteduniverse of values
of D, P and R. Specifying the function in extension merely tells us that
whenever the system is given some D and P as input, it should yield some
specifiedR as output. One way of doing this (and I here adapt a strategy used by
Martin Davies (see Davies [forthcoming] and Section 3 following), is to store
the set of legal values of R for every combination of values of D and P-a simple
look-up table. A second way is to process data in accordance with the equation
P = D x R. In saying that the system draws on the information that P = D x R,
we are, as Peacocke insists, doing morethan specifying a function in extension.
For the look-up table does not draw on that information, yet it falls within the
equivalence class generated by the function in extension specification. But we
are doing less than specifying a particular algorithm, since there will be many
ways of computing the equation in question (e.g. using differentalgorithms for
multiplication).
It is this grain of analysis (i.e. what Peacocke calls level 1 5) that I will have
in mind when speaking, in the remainder of this paper, of a competencetheory.
This seems to accord at least with the practice of Chomsky, who coined the
term competence theory to describe the pitch of his own distinctive
investigations into the structure of linguistic knowledge. And it may well

200 Andy Clark
accord with Marr's actual practice at 'level 1' though not with the official
dogma.
A Chomskian competence theory does far more than specify a function in
extension. Instead it seeks to answer (at a level of abstraction from the physical
mechanisms of the brain and from specific algorithms) the question 'What
constitutes knowledge of language?'. In so doing it seeks a 'framework of
principles and elements common to attainable human languages' (Chomsky
[1986], p. 3). And (in its most recent incarnation) it characterizes that
framework as a quite specific
system of principlesassociatedwith certain parametersof variation and a
markednesssystem with severalcomponentsof its own (Chomsky[1986], p.
221).
It does not matter, for the purposes of this paper, just what principles and
parameters Chomsky actually suggests. Rather, we should merely note that if
a competence theory is as definite and structured as a Chomskian model (and
its his word, after all!) then it is more like a level 5 analysis than a simple level
1 account. For it describes, at a certain level of1.abstraction, the structure of a
form of processing (by specifying the information drawn on by the processes)
and hence helps 'guide the search for mechanisms' (Chomsky [1986], p. 221).
In short, it is more like Mendelian genetics than General Darwinism. Rather
than being merely descriptiveof a class of results, it is meant also to be suggestive
of the processing structure of a class of mechanisms of which we are a member.
Whether competence theories (at least as we currently know them) actually
are suggestive of the form of human processing is the topic of this paper.
3 THE CLASSICAL CASCADE
A competence theory, then, leads a double life. It both specifies the function to
be computed and it specifies the body of knowledge or information which is
used by some class of algorithms. In classical cognitive science, these two roles
can easily be simultaneously discharged. For the competence theory is just an
articulated set of rules and principles defined over symbolic data-structures.
Since classical cognitive science relies on symbol processing architecture, it is
natural (at level 2) to represent directly the data structures (e.g. structural
descriptions of sentences) and then carry out the processing by the explicit or
tacit representation of the rules and principles defined (in the competence
theory) to operate on those structures. Thus, given a structural description of
an inflected verb as comprising a stem plus an ending, the classicist can go on
to define a level 2 computational process to take the stem and add -ed to form
the past tense (or whatever). The classicist, then, is (by virtue of using a symbol
processing architecture to implement level 2 algorithms) uniquely well placed
to preserve a very close relation between a competence theory and its level 2

Connectionism, Competence, and Explanation 201
implementations. Indeed, it begins to seem as if that close relation is what is

constitutive of a classical approach. Thus Dennett visualizes the classicist dream
as involving 'a triumphant cascade through Marr's three levels' (Dennett
[ 19 8 7], p. 22 7). Such a characterization of the essential classicist vision seems
to me to fit very nicely with Fodor and Pylyshyn's recent account of the
classical/connectionist divide.
Fodor and Pylyshyn argue that there are two fundamental differences
between truly connectionist and classical approaches to cognitive modelling.
('Truly connectionist' here rules out those cases where a units and connec-
tions sub-structure is used to implement a classical theory). The differences
are:
1. 'Classical theories-but not connectionist theories-posit a "language of
thought"'.
This means they posit mental representations (data-structures) with a
certain form. Such representations are syntactically structured,i.e. they are
systematically built by combining atomic constituents into molecular
assemblies which in turn (in complex cases) make up whole data-structures.
In short, they posit symbol systems with a combinatorial syntax and
semantics.
2. 'In classical models, the principlesby which mental states are transformed,or
by which an input selects the corresponding output, are defined over
structural properties of mental representations. Because classical mental
representationshave combinatorial structure, it is possible for classical mental
operationsto apply to them by reference to their form'.
This means that given that you have a certain (language-like) kind of
structured representation available (as demanded by point 1), it is possible to
define computational operations on those representations so as the oper-
ations are sensitive to that very structure. If the structure was not there (i.e. if
there was no symbolic representation) you could not do it! (Though you
might make it look as if you had by fixing a suitable function in extension.)
(Quotes are from Fodor and Pylyshyn [1988], pp. 12-13.)
In short, a classical system is one which posits syntactically structured
symbolic representations and which defines its computational operations to
apply to such representations in virtue of their structure.
The computational operations, in any such case, can be described by
transition or derivation rules defined over syntactically structured represen-
tations. For example:
If (A and B) then (A)

If (A and B) then (B)
If (stem + ending) then (stem + -ed)
The parenthesized items are structural descriptions which will pick out open-
ended classes of classical representations. The 'If-then' specifies the operation.

202 Andy Clark
But note that the classicist, under the terms of the act, is not committed to the
systems explicitlyrepresenting the 'if-then' clause. All that needs to be explicit
is the structured description upon which it operates. Thus a machine could be
hard-wiredso as to take expressions of the form (A and B) and transformthem
into the expressions (A) and (B). The derivation rules may thus be implicit, or
tacit; but the data-structures must be explicit. On this matter, Fodor and
Pylyshyn are rightly insistent:
Classicalmachinescan be ruleimplicitwith respectto theirprograms.... What
doesneedto be explicitin a classicalmachineis not its programbut the symbols
that it writeson its tapes(or storesin its registers).These,however,correspond
not to the machine'srulesof statetransitionbutto its datastructures.(Fodorand
Pylyshyn[1988] p. 61.)
As an example they point out that the grammar posited by a linguistic theory
need not be explicitly represented in a classical machine. But the structural
descriptionsof sentencesover which the grammar is defined (e.g. in terms of verb
stems, sub-clauses etc.) must be. A successful 'classical cascade' from a
linguistic competence theory to a level 2 processing story can thus tolerate
having the rules of the grammar built into the machine. Those attempts to
characterize the classicist/connectionist contrast solely by reference to the
explicitness or otherwise of rules are thus shown to be in error.
Now, however, there is a danger of losing sight of the way in which (for a
classicist) the competence theory (or set of derivation rules and data-
structures) is meant to bear a close relation to the level 2 implementation. For
we said (Section 2 above) that given, say, a simple competence theory like
'P = D x R' it wouldnot do to have a system which simply stored, for some finite
universe of discourse, all legal values of P, D, and R. Yet such a system certainly
has explicit representations of P, D, and R. So if it will not do, it must be because
it lacks even tacit knowledge of the derivation rule 'P = D x R'. The question
then is, how do we motivate this difference?What are the constraints on tacit
knowledge ascription such that the rule 'P=D x R' need not be explicitly
represented, but which rules out the look-up table as an instance of tacit
knowledge of that very rule?The answer will be significant when we come to
ask (Sections 4 and 5) whether connectionist systems have tacit knowledge of
classical rules.
MartinDavies (drawing on information providedby GarethEvans) offersthe
following suggestion:
Fora Speakerto have tacit knowledgeof a particulararticulatedtheory,there
must be a causal-explanatorystructurein the Speakerwhich mirrorsthe
derivationalstructurein the theory(Davies[forthcoming],p. 4).
By 'the derivational structure in the theory' Davies means the transition rules
(e.g. P=D x R). What is it, then, to embody a 'causal-explanatory structure'
which 'mirrors'such a derivational structure?Simply, according to Davies, for

there to be a causalcommonfactor (Davies' phrase) in the processing story told
for each instance which, at the higher level, is seen as involving the rule of
derivation. Thus, in the case of the look-up table, there need be no causal
common factor in the processing of all the instances of various values of P, D,
and R. Conversely, if there is a causal common factor through which all
processing is routed, then (subject to some niggling provisos-see Davies
[forthcoming] and [1987]) the system is rightly said to have tacit knowledge of
the rule. A result which, as Davies notes, sits nicely with our cognitive
neuropsychological intuitions. For systems which meet the tacit knowledge
constraint so construed are prima facie candidates for a type of breakdown in
which damage to the causal common factor causes total loss of capacity to
solve a whole class of problems (e.g. P, Q, and R specifying). Whereas systems
which fail to meet the constraint are prima facie candidates for less systematic
deficits-e.g. the look-up system may lose its knowledge of some legal
combinations of P, Q, and R but preserve its knowledge of others. Similar
comments apply to the past tense generation case. Systems with tacit
knowledge of the rule 'take stem and add -ed' could lose all capacity to form
regular pasts. Systems which do it by look-up would not.
Davies' account (modulo a quibble about virtual'causal' common factors see
Section 4 following) seems convincing. If we take it on board, we end up with
the following characterization of any properly classicalcognitive model:
(Classicismdefinedby an attitudeto competencetheories.)A cognitivemodelis
classicaliffit has a processingleveldescriptionwhichbearsa certainratherclose
relation to the structure of a standard competence theory. A standard
competencetheorypositsa set of rulesorprinciplesof derivationdefinedto apply
to a class of structured,symbolicrepresentationsaccordingto theirform.The
closerelationrequiredinvolves(1) the explicitrepresentation, in the processing
level description,of the structuredrepresentationsover which the rules are
defined,andit involves(2) the explicitORtacitrepresentation of thoserulesand
principlesthemselves.A ruleorprincipleis judgedto be tacitlyrepresented justin
case thereis a causalcommonfactorin the processingleveldescriptionwhichis
in play whenever the rule or principle is invoked in a competence-level
specificationof a transition.
Such, in tortuous detail (and apologies for that) is the substance of the 'classical
cascade' through Marr's levels of explanation. Connectionism dams the
cascade. How it does so, and what water-courses result, will occupy us for the
remainder of this paper.
4 NEWTONIAN COMPETENCE
The connectionist vision of the relation between a structured competence

theory and a level 2 processing story is radically unlike the neat 'cascade'
imagined in Section 3. Instead of the level 2 story mirroring the derivational

204 Andy Clark
form of the competence theory, it is seen as relating to it in rather the way
Newtonian mechanics relates to quantum physics. The physical universe is
not, in fact, Newtonian. But under certain specifiable conditions, it behaves
very must as if it were. Newtonian principles thus describe and predict the
behaviour of physical systems in a range of cases. But, in some intuitive but
slightly elusive sense, those principles do not describe the actual forces which
determine physical behaviour. The analogy is much favoured by Rumelhart
and McClellandwho write:
It mightbe arguedthatconventionalsymbolprocessingmodelsaremacroscopic
accounts,analogousto Newtonianmechanics,whereasour modelsoffermore
microscopicaccounts, analogousto quantum theory-Through a thorough
understandingof the relationshipbetween the Newtonian mechanics and
quantumtheorywe can understandthat the macroscopiclevel of description
may be onlyan approximation
to the moremicroscopictheory.(Rumelhartand
McClelland[1986] p. 125).
To illustrate this point, consider a simple example due to Paul Smolensky.

Imagine that the cognitive task to be modelled involves answering qualitative
questions concerning the behaviour of a particular electrical circuit. (The
restriction to a single circuit may appal classicists, although it is defended by
Smolensky on the grounds that a small number of such representations may
act as the 'chunks' utilized in general purpose expertise-see Smolensky
(1986] p. 241).) Given a description of the circuit, an expert can answer
questions such as 'Ifwe increase the resistance at point A what effect will that
have on the voltage?' (i.e. will the voltage increase, decrease, or remain the
same?).
Suppose, as seems likely, that a high-level competence-theoretic specifica-
tion of the information to be drawn on by an algorithm tailored to this task
cites various laws of circuitry on its derivations (what Smolensky refers to as
the 'hard laws' of circuitry; Ohm's law and Kirchoff's law). For example,
derivations involving Ohm's law would invoke the equation
Voltage (V)= Current (C)x Resistance (R).
We recognized, in Section 3 above, just two ways in which a level 2 processing
story might bear an appropriatelyclose relation to such a competence theory.
In the simplest case, the processing might involve a symbolic representationof
Ohm's law which is read and followed by the system. In the more complex case,
it might involve tacit knowledge of Ohm's law unpacked in terms of a causal
common factor in a set of state transitions. (Note in passing: Smolensky's own
treatment here seems to place uncalled for emphasis on the simple option-see
Fodor and Pylyshyn [1988], Pinker and Prince [1988], Davies [forthcoming],
Clark [1989].)

Neither cascade is operative in the case of Smolensky's connectionist model
of simple circuit problemsolving. To see why, we need to look at the form of the
model in question.
The model representsthe state of the circuit by a pattern of activity over a set
of feature units. These encode the qualitative changes found in the circuit
variables, i.e. in training instances, they encode whether when the resistance
at R1 goes up, the overall voltage falls or rises and so forth. These feature units
are connected to a set of what Smolensky calls 'knowledge atoms' which
represent patterns of activity across subsets of the feature units. These in fact
encode the legal combinations of feature unit states allowed by the actual laws
of circuitry. Thus, for example:
The system's knowledge of Ohm's law.., .is distributedover the many
knowledgeatomswhose subpatternsencodethe legalfeaturecombinationsfor
current,voltageand resistance.(Smolensky[1988] p. 19.)
In short, there is a sub-pattern for every legal combination of qualitative
changes (GS sub-patterns, or 'knowledge atoms' for the circuit in question).
It might seem, at first sight, that the system is merely a units and
connections implementation of a look-up table. But this is not so. In fact,
connectionist networks act as look-up tables only when they are providedwith
an overabundance of hidden units and hence simply memorize input-output
pairings. By contrast, the system in question encodes what Smolensky terms
'soft constraints', i.e. patterns of relations which usually obtain between the
various feature units (microfeatures). It thus has 'general knowledge' of
qualitative relations among circuit microfeatures. But it does not have the
general knowledge encapsulated in hardconstraints like Ohm's law. The soft
constraints are two-way connections between feature units and knowledge
atoms which inclinethe network one way or another, but do not compelit; that
is, they can be overwhelmed by the activity of other units-that is why they
are 'soft'. And as in all connectionist networks, the system computes by trying
simultaneously to satisfy as many of these soft constraints as it can. To see that
it is not a mere look-up tree of legal combinations we need only note that it is
capable of giving sensible answers to (inconsistent or incomplete) questions
which have no answer in a simple look-up table of legal combinations.
The soft constraints are numerically encoded as weighted inter-unit
connection strengths. Thus problem solving is achieved by 'a series of many
node (i.e. unit) updates, each of which is a microdecisionbased on formal
numericalrules and numerical computations' (Smolensky [1986], p. 246).
The network has two properties of special interest to us. First, it can be
shown that if it is given a well-posed problem andunlimited processing time it
will always give the correct answer as predicted by the hard laws of circuitry.
But, as already remarked, it is by no means bound by such laws. Give it an ill-

206 AndyClark
posedor inconsistentproblemandit willsatisfyas manyof the softconstraints
(which are all it reallyknows about)as it can. Thus:
Outsideoftheidealized
domainofwell-posed andunlimited
problems processing
time,thesystemgivessensibleperformance [1988],p. 19).
(Smolensky
The hardrules(Ohm'slaw etc.) can thus be viewedas an externaltheorist's
characterizationof an idealizedsubset of its actual performance(it is no
accident if this puts us in mind of Dennett's [1987] claims about the
'intentionalstance').
Second,the networkexhibitsinterestingserialbehaviouras it repeatedly
triesto satisfyall the softconstraints.Thisserialbehaviouris characterized
by
Smolenskyas a set of macrodecisions each of which amountsto a 'commitment
of part of the networkto a portionof the solution'.These macrodecisions,
Smolenskynotes, are:
approximatelylikethefiringofproduction
rules.Infact,these'productions'
'fire'
inessentially
thesameorderasina symbolic forward-chaining inference
system
(Smolensky[1988],p. 19).
Thusthe networkwill lookas if it is sensitiveto hard,symbolicrulesat quitea
fine grain of description.It will not simplybe that it solves the problem'in
extension'as if it knewhardrules.Eventhe stagesof problemsolvingmay look
as iftheyarecausedby the system'srunninga processinganalogueofthe steps
in the symbolicderivationsavailablein the competencetheory.
But the appearanceis, on the termsset out in Section3 above,an illusion.
Thesystemhas neitherexplicitnor tacitknowledgeof the hardrules.It is not
hard to see why. Quiteclearly,it does not explicitlyrepresentOhm'slaw to
itself.Thereis, forexample,no neat sub-patternof units whichcan be seen to
standfor the generalidea of Resistancewhich figuresin Ohm'slaw. Instead,
sets of units standforResistance-at-RI, and othersets forResistance-at-R2.
In
morecomplexnetworks,the coalitionsof units which, when active,standin
for a top (or conceptual)level concept like resistanceare highly context-
sensitive.Thatis, they vary accordingto context of occurrence.Thus,to use
Smolensky'sown example,the representationof coffeein such a network
wouldnot comprisea singlerecurrentsyntacticitembut a coalitionof smaller
items(microfeatures) whichshiftaccordingto context.Coffeein the contextof
cup may be representedby a coalitionwhich includes(liquid)(contacting-
porcelain).Coffeein the context of jar may include (granule)(contacting-
glass).Thereis thus only an 'approximateequivalenceof the "coffeevectors"
across contexts' unlike the 'exact equivalenceof the coffee tokens across
differentcontextsin a symbolicprocessingsystem'(Smolensky[1988], p. 17).
By thus replacingthe conceptuallevelsymbol'coffee'with a shiftingcoalition
of microfeatures, the so-called 'dimension shift', such systems deprive

themselves of the structured mental representations which are deployed both
in a classical competencetheory and in a classical symbol processing (level 2)
account. Likewise, there is no stable representational entity in the simple
network described which stands for Resistance (just as in the infamous past-
tense network there is no stable, recurrent entity which stands for 'verb-stem'
(see Rumelhart and McClelland [1986b], Pinker and Prince [1988], Clark
[1989]). The immediate result is that there can be no explicit representation of
rules which involve reference to the conceptual level constructs. The lack of
tacit representation is almost immediate, since the processing can hardly be
sensitive to structures which are not there.
To put the point in our favoured terms, the system cannot be said tacitly to
represent the rules since there is no causal common factor in its problem
solving such that whenever, e.g., Ohm's law would be cited in the competence
theory, that single factor is pivotal in the processing which yields the actual
result. To see this we need only reflect that different feature units and
knowledge atoms will be pivotal in solving problems which relate to the fate of
R1and ones which relate to the fate of R2. In this (restricted)sense it doeshave
something in common with the look-up tree. For the network fails to embody
strict tacit knowledge of the rule because it fails to route all its actual
processing through a causal bottleneck corresponding to the derivational
bottleneck marked by the repeated citing of Ohm's law. By having multiple
causal routes where the competence theory has a single derivational equation,
the network loses its claim to strict tacit knowledge of the rule. In that respect,
it fails to embody tacit knowledge of the rule for the same reason as does the
look-up tree.
Now for the quibble promised earlier. In adopting, as far as I understand it,
Davies' characterization of tacit knowledge, I am uneasy about the use of the
phrase 'causal common factor'. It has the advantage of making neuropsycho-
logical implications seem very immediate. But it may paper over some of the
complexities of stacked virtual machines. For my guess is that what would
need to be common for the classical cascade to be realized, is not a simple
physical state so must as a state of the virtualmachineover which theprocessing
story is defined.After all, even a classical system, courtesy of various niceties of
operating systems, may not use the same physical state every time it goes
through a processing transition marked (in the competence theory) by Ohm's
law. However, the level 2 processing description need not (and ought not)
signal the difference,since it has no implications as far as the actual algorithm
is concerned. It is merely an implementation detail. Contrariwise,the variety of
states which, in a connectionist story, may correspond to a single symbolic
transition, must be signalled in the processing/algorithmic description. After
all, the system's real knowledge is the knowledge so encoded-a fact which is
directly responsible for the much-vaunted fluidity and context-sensitivity of
connectionist processing. I am not sure how much of a differencethis makes

208 Andy Clark
since virtual machines, as much as real ones, can exhibit distinctive
breakdown patterns and hence tie in with the cognitive neuropsychology.
Quibblingaside, we are now in a position to sum up the Newtonian attitude
to competence theorizing. A Newtonian connectionist will regard a com-
petence theory as descriptive (perhaps at a quite fine grain-recall the
discussion of 'macrodecisions') of the course of processing. But she will not
regard it as suggestiveof the actual processing involved. It is not suggestive
because the behaviour is not dependent on the system's having explicit or tacit
knowledge of the symbolic derivation rules; a fact evidenced in its behaviour
outside the idealized, 'Newtonian' domain of well-posed problems and
unlimited processing time. This behaviour shows that 'it's really been a
"quantum" system all along' (Smolensky [1988], p. 19).
In a revealing footnote (Smolensky [1986], p. 246) the point is cast in terms
highly appropriateto our discussion. The characterization of competence as a
set of derivation rules applied to a symbol system can be viewed, Smolensky
suggests, as providing a grammarfor generating the high-harmony (= maxi-
mal soft constraint satisfaction) states of a system. Thus a competence theory
emerges as a body of laws which serve to pick out the states into which the
system will settle in certain ideal conditions. This, then, is the full Newtonian
attitude to a competence theory: a competence theory is a kind of grammar
which fixes on certain stable states of the system. As such it is, in a central
range of cases, descriptively adequate. But it does not reveal what Smolensky
calls the dynamics,or actual processing strategies, of the system. It is not a
properly suggestive guide to the level 2 processing story. For the Newtonian
then, competence theorizing just ain't what it used to be.
5 ROGUE COMPETENCE
On the Newtonian connectionist model, then, the competence theory
functions as a descriptively adequate guide to the output in a somewhat
idealized range of cases. This, however, is not the only understanding of
competence theories available to a connectionist. And indeed, it is not the
understanding implicit in some other connectionist treatments of high level
problem solving. In this Section I look at a class of alternative treatments
which I shall call roguemodels of competence.
The basic differencebetween Newtonian and rogue models is simply this. In
a Newtonian model, the connectionist network is itselfcapable, under idealized
conditions, of behaving in all the ways specifiedby the competence theory. In a
rogue model, by contrast, the basic connectionist network does not itself have
the capacity (even under idealizations of processing time and well-posed
problems) to produce the full range of results requiredby (i.e. derivable in) the
competence theory. Instead, it will be claimed that insofar as human beings
actually exhibit the full scale classical competence they do so only by deploying

other resources (for example, a linked symbol processor or real world
structures (like pen and paper) for manipulating symbols). The view of
competence models which emerges from a rogue approach is thus that they
involve pressing into service extra resources which are not on-line in fast daily
problem-solving in the domain.
An example of a rogue model can be found in Rumelhart, Smolensky,
McClelland, and Hinton [1986]. The example concerns our capacity to
multiply numbers. We might imagine a symbolic competence theory here
appealing to the laws of arithmetic. But a basic connectionist model will not
resemble such a symbolic store. Rather, it will amount to a well trained pattern
matcher which can 'see' the results of some multiplications right away. For
example, most of us can 'see' the answer to 7 x 7, but not to 7984 x 5431.
How then, do we solve the latter kind of problem?
The conjecture is that:
The answer comes from our abilityto create artifacts-that is, our abilityto
createphysicalrepresentationsthat we can manipulatein simpleways to get
answers to very difficult and abstract problems.(Rumelhart,Smolensky,
McClelland, and Hinton[1986] pp. 44).
Thus, to solve 7984 x 5431 we might write down the question and then solve
it by the careful deployment of a series of the simple pattern-matching steps we
are good at, e.g. beginning by multiplying 4 x 1 and so on:
7984
5431
4.
We may, they go on to say, even learn to do this in our headby representing the
external symbols to ourselves in some manner. But it is still an essentially
'external' symbolic medium which we are manipulating, and it still constitutes
a resource built on top of the basic connectionist pattern-matching capacity
which we deploy. (Daniel Dennett has recently being saying very similar
things about the cases where sentencesseem to run through our heads. In these
case, we do indeed do classical symbol processing. But such processing may
constitute an extra resource, not implicated in all our daily, non-linguistic
reasoning-see Dennett [1987], pp. 233, 114-15; also Clark [1988].)
The account of complex multiplication is of course highly problematic since
the whole thing seems to involve knowing symbolic rules governing the serial
deployment of the pattern-matching capacities! But we have seen already that
much apparentlysymbol-reliant behaviour may be sub-symbolically produced.
(But see Clark [1989] for a detailed discussion.) And at any rate, I use the
example merely as a gesture at the kindof account which would constitute a
rogue model.

2IO Andy Clark
To give one final example (which I owe to Martin Davies), consider our
capacity to parse garden-path sentences like:
the horse raced past the barn fell.
A rogue model of parsing might go something like this. We have on-line a
quick and dirty connectionist network which can parse most of the sentences
we encounter in daily speech. But it does not have the capacity (even in
principle, subject to idealization) to parse a garden-path sentence. However,
we also have (not on-line, but in the background) a classical symbolic parser
(something like an ATN?) which can parse such cases. And when the quick
and dirty network fails, this back-up comes on-line to save the day. This fits the
phenomenology, in which the sentence at first looks like nonsense, then falls
into place. In such a case the classical competence theory correctly describes
the structure of the back-upsystem. But it does not describethe on-line network.
If, in addition, we imagine that the classical back-up system was active in
training up the network, the partial confluence of the two systems over a range
of simple cases is rendered unsurprising.
An obvious and related advantage of the rogue approach concerns the
psychological plausibility of so-called supervised learning algorithms. These
are procedures for training connectionist networks which rely on the back
propagation of error messages, and hence rely on a teacher (usually a
conventional computer) which looks at the system's output and tells it what
the output shouldhave been like. (For a little more detail, see the discussion of
NETtalkin Section 6.) Such set-ups have often appeareddeeply psychologically
unrealistic. For example, when we learn a language, we can do so by being
given positive examples only (as Chomskians are fond of pointing out).
Whence, then, the teacher and the error messages?
The possibilitywhich rogue models open up is that a separate system stores a
set of input-output pairings (e.g. a set of observed print-phoneme pairings) and
uses these to train a connectionist network. The negative instances are thus
generated and spotted by the brain itself, rather than by other agents. Terence
Sejnowski has recently endorsed such a picture and illustrates it by citing the
case of the White Crown Sparrow which hears its father's song one year but
does not sing it until the next. The hypothesis is that the bird somehow stores
the song, but must train up a network to reproduce it-a process which
explains the long gap between exposure and reproduction. White Crown
Sparrows aside, rogue approaches clearly offer the best hope for the
psychological respectability of the back propagation method of connectionist
learning.
At its most extreme, a rogue model may divorce human on-line processing
from the strict competence model, but reinstatethe classical competence as a
full and proper description of a back-up system. Note that the status of the
classical competence theory on a rogue model is quite differentfrom its status

on a Newtonian one. For the rogue modeller, the classical competence theory
properly describes an important, though not constantly on-line, class of
processing systems. In fact, the importance of these classical resources is, I
suspect, not yet fully appreciated even where lip service is paid to their
presence. Thus Smolensky [forthcoming] introduces the idea of language as a
special medium of knowledge transmission involving processing by a classical
virtual machine called the Conscious Rule Interpreter.But the role of linguistic
instruction is still presented as somewhat second-grade. Language allows us to
formulate rules which, for example, help the novice in the early stages of
training (see also Smolensky [1986], pp. 251-2, where essentially the same
picture is applied to the previously discussed case of electric circuit problem
solving). The expert, however, is pictured as using a powerful connectionist
network, and seems to need language only to transmit potted elements of her
insights to others. This may severely under-estimate the contribution of
symbol processing. Such processing may also help the expert to understand
and extend her own skills by providing a kind of meta-reflection on her own
on-line reasoning. (For some related hypotheses see Karmiloff-Smith[1987]
and Dennett [1988].)
The most potent effect of the adoption of a rogue approach is vastly to
complicate the currently fashionable debate concerning the 'correct'cognitive
architecture of mind (see Fodor and Pylyshyn [1988]). For if a rogue model is
adopted, there is no unique answer to such questions. Any good account of
human cognitive skills will need to employ both kinds of model, and the
classical version will not be just a convenient approximation. It is as if the
physical world turned out to be Newtonian in some areas and quantum in
others, rather than being uniformly quantum-describable but in some
circumstances lookingNewtonian.
To sum up, rogue models deny even the descriptiveadequacy of classical
competence models to on-line processing. But they allow that the classical
theory is both descriptive and suggestive of the processing of an additional
resourcesystem. This additional resource system guarantees what might be
called our canonicalreasoning abilities in a given domain. In rogue cases, the
competence model is what it used to be (an accurate description of some
processing strategy), but it is not whereit used to be-for it does not describe
the computational form of daily on-line processing.
6 THE METHODOLOGY OF CONNECTIONIST EXPLANATION
Connectionist explanatory strategies, it seems, cannot fit into the mould

suggested by Newell and Simon. A connectionist cannot begin with a Newell
and Simon style competence theory and then simply implement it in a level 2
algorithmic model. The reason, we saw, is straightforward.Such a competence
theory consists of a set of transition rules defined to apply to standard symbolic

212 Andy Clark
representationsor data-structures. In a classicalmodel,thesedata-structures
areexplicitlyrepresentedin the machine(classicalfunctionalarchitecturesare
preciselythosearchitectureswhichmakethis possible).Andthe machinethen
manipulatesthemin accordancewith the rules(whichneednot themselves be
explicitlytokenedin any suchdata-structures). In a distinctivelyconnectionist
model,by contrast,there will be nothing which neatly correspondsto the
classicalsymbolicdata-structures.Instead,context-sensitive,shiftingcoali-
tions of units will correspondto single classicalrepresentations.This is the
dimension-shift describedearlier.Sincetherearethusno neatanaloguesto the
classical symbolicstructures,the system cannot(not even tacitly) embody
knowledgeof transitionrulesdefinedover thoseverystructures. So a classical
competencetheory cannot be richly suggestive of a connectionist level 2
processingstory. If it were, then the 'connectionist' system would amount
merelyto a fast, robust,implementationof a classicalcognitivemodel (see
Fodorand Pylyshyn[1988]).
Givenall this, we saw that the devoutconnectionistcouldadoptone of two
positionsregardingthe classicalcompetencemodel.Thesewere the Newto-
nian and Roguepositionsdiscussedabove. But a deeper,foundationalissue
remainsunresolved.For the Newtonianand Rogue positionsare united in
denyingthatany toplevelclassicalcompetencetheorycanberichlysuggestive
ofthe level2 processingstrategiesofthe centralon-lineconnectionistnetwork
which carriesout a givencognitivetask.Butthis (recallSection2 above)now
looks to be a doublyembarrassingloss. Forthe classicalcompetencetheory
performedtwo tasks. First, it figured in a picture of the properform of
investigationsin cognitive science (i.e. delineate the task at the level of
competencetheorizingandthen writealgorithmsto carryit out).Andsecond,
it figuredin a pictureof what explanation in cognitivescienceinvolved.Just
a
having workingprogram was not, in itself, to be regardedas having an
explanation of how we a
perform given cognitive task. Rather,we wanted
somehigh-levelunderstandingof what constraintsthe programwas meeting
and why they had to be met-an understandingnaturallyprovidedby giving
the toplevelcompetencetheorywhich a givenclassof programscouldbe seen
to implement.The unavailabilityof the classical competencetheory thus
threatensto renderconnectionistmodelsnon-explanatory in a verydeepsense.
And it leaves the actual methodology of connectionistinvestigationsobscure.
As a briefillustrationof the problem,consideran exampleof GoodOld
FashionedExplanationIn CognitiveScience (GOFEICS-apologies to John
Haugeland).TakeNaive Physics.Naive Physics, as everyoneknows, is the
attemptto discoverthe knowledgewhich enablesa mobile,embodiedbeingto
negotiateits way arounda complexphysicaluniverse.A well-knowninstance
of this generalprojectis Hayes'[1984] workon the naive physicsof liquids.
Thisinvolvedtryingto compilea 'taxonomyof the possiblestatesliquidcan be
in' and formulating a set of rules concerning movement, change and liquid

geometry. The final theory included specifications of fifteen states of liquid and
74 numbered rules or axioms written out in predicate calculus. This amounts
to a detailed competence specification which might eventually be given full
level 2 algorithmic form. Indeed, Hayes ([1985], p. 3) is quite explicit about the
high level of the investigative project, insisting that it is a mistake to seek a
working program too soon. The explanatory strategy of naive physics is thus a
paradigm example of the official classical methodology recommended by
Newell and Simon. First, seek a high level competence theory involving
symbolic representations and a set of state transition rules. Then write level 2
algorithms implementing the competence theory, secure in the knowledge
that we have a precise higher level understanding of the requirements which
the algorithms meet and hence a real grasp of why they are capable of carrying
out the task in question. It is this security which the connectionist lacks, since
she does not (cannot)proceed by formulating a detailed classical competence
theory and then neatly implementing it on a classical symbol processing
architecture.
Hence the problem: how should the connectionist proceed, and what
constitutes the higher level understanding of the processing which we need in
orderto claim to have really explainedhow a task is performed?What is needed,
it seems, is some kind of connectionist analogue to the classical competence
theoretic level of explanation.
I believe that such an analogue exists. But it remains invisible until we
perform a kind of Copernican revolution in our picture of explanation in
Cognitive Science. For the connectionist effectively inverts the usual temporal
and methodological order of explanation, much as Copernicus inverted the
usual astronomical model of the day by having the earth revolve around the
sun instead of the other way round. Likewise, in connectionist theorizing,
the high level understanding will be made to revolve around a working
program which has learnt how to negotiate some cognitive terrain. This
inverts the official Marr-style ordering in which the high level understanding
(i.e. competence theory) comes first and closely guides the search for
algorithms. To make this clear, and to see how the connectionist's high level
theory will depart from the form of a classical competence theory, I propose to
take a look at Sejnowski's NETtalkproject.
NETtalkis a large, distributedconnectionist model which aims to investigate
part of the process of turning written input (i.e. words) into phonemic output
(i.e. sounds or speech). The network architecture comprises a set of input units
which are stimulated by seven letters of text at a time, a set of hidden units, and
a set of output units which code for phonemes. The output is fed into a voice
synthesizer which produces the actual speech sounds.
The network began with a random distribution of hidden unit weights and
connections (within chosen parameters), i.e. it had no 'idea' of any rules of text
to phoneme conversion. Its task was to learn, by repeated exposure to training

214 Andy Clark
instances, to negotiate its way around this particularly tricky cognitive

domain (tricky because of irregularities, sub-regularities and context-sensiti-
vity of text-+phoneme conversion). And learning proceeded in the standard
way, i.e. by a back-propagationlearning rule. This works by giving the system
an input, checking (this is done automatically by a computerized 'supervisor')
its output, and telling it what output (i.e what phonemic code) it shouldhave
produced. The learning rule then causes the system to minutely adjust the
weights on the hidden units in a way which would tend towards the correct
output. This procedure is repeated many thousands of times. Uncannily, the
system slowly and audibly learns to pronounce English text, moving from
babble to half-recognizable words and on to a highly creditable final
performance. (For a full account, see Rosenberg and Sejnowski [1987] and
Sejnowski and Rosenberg [1986].)
Considernow the methodology of the NETtalkproject. It begins, to be sure,
by invoking the results of some fairly rich prior analysis of the domain. This is
reflected in the author's choice of input representation (e.g. the choice of a
seven letter window, and a certain coding for letters and punctuation), in the
choice of output representation (the coding for phonemes) and in the choice of
hidden unit architecture (e.g. the number of hidden units) and learning rule.
These choices highlight the continued importance of some degree of prior task
analysis in connectionist modelling. But they are a far cry from any fully
articulated competence theory of text to phoneme conversion. For what is
noticeably lacking is any set of special purpose state transition rules defined
over the input and output representations. Instead, the system will be set the
task of learning a set of weights over its hidden units such that the weights
perform the task of mediating the desired state transitions. For this reason I
shall characterize the connectionist as beginning her investigations with a
level 5 'task analysis', as opposed to a level 1 (or 1 5) competence theory. It is
worth? remarking, however, that the level specification, though less than a
?5 still
full-blown symbolic competence theory, may embody a psychologically
unrealistic amount of prior information. For when a human learns to perform
a task she does not know, in advance, how many hidden units to allocate (too
many and you form an uninformative 'look-up tree', too few and you fail to
deal with the data) or the best way to represent the solution. In this sense, the
level -5 specification may be doing more of the problem solving work than
some connectionists would like to admit. For present purposes, however, the
point is just that the level *5model forms the basis upon which, courtesy of the
powerful connectionist learning rules, the system comes to be able (aftermuch
training) to negotiate the targetted cognitive terrain. At this point, the
connectionist has in her hand a working system-a full-scale level 3
implementation.
Suppose we were to stop there. We would have a useful toy, but very little in
the way of increased understanding of the phenomenon of text-phoneme

conversion. But, of course, the connectionist does not stop there. From the up
and running level 3 implementation she must now work backwardsto a
higher-level understanding of the task. This is Marr-through-the-looking-
glass. How is this higher level understanding to be obtained? There are a
variety of strategies in use and many more to be discovered. I shall mention just
three. First, there is simple watching, but at a microscopic level. Given a
particular input, the connectionist can see the patterns of unit activity (in the
hidden units) which result. (This, at any rate, will be the case if the network is
simulated on a conventional machine which can keep a record of such
activity). This, as Sejnowski points out, provides a kind of data which
neuroscientists are hard pressed to gather. For neuroscience has excellent
techniques for recording single cell activity. But it is not well placed to record
patterns of simultaneous activity across large numbers of cells. (See also
Churchland [forthcoming-1989].)
Second, there is networkpathology.While it is obviously unethical delibera-
tely to damage human brains to help us see what role sub-assemblies of cells
play in various tasks, it seems far more acceptable to damage artificial neural
networks.
Lastly, and perhaps most significantly, the connectionist can generate a
picture of the way in which the system has learnt to divide up the cognitive
space it is trying to negotiate. It is this picture, given by so-called 'hierarchical
cluster analysis', which seems to me to offerthe closest connectionist analogue
to a high-level, competence-theoretic understanding.
Cluster analysis is an attempt to answer the question, 'What kinds of
representation have become encoded in the network's hidden units?' This is a
hard question since the representations, as noted earlier, will in general be of
somewhat complex, unobvious, dimension-shifted features. To see how cluster
analysis works, consider the task of the network to be that of setting hidden
unit weights in a way which will enable it to performa kind of set partitioning.
The goal is for the hidden units to respond in distinctive ways when, and only
when, the input is such as to deserve a distinctive output. Thus in text-to-
phoneme conversion, we want the hidden units to perform very differently
when given 'the' as input than they would if given 'sail' as input. But we want
them to perform identicallyif given 'sail' and 'sale' as inputs. So the hidden
units' task is to partition a space (definedby the number of such units and their
possible levels of activation) in a way which is geared to the job in hand. A very
simple system, such as the rock/mine network described in Churchland
[forthcoming-1989] may need only to partition the space defined by its
hidden units into two major subvolumes-one distinctive pattern for inputs
signifying mines and one for those signifying rocks. The complexities of text-
phoneme conversion being what they are, NETtalkmust partition its hidden
unit space more subtly (in fact, into a distinctive pattern for each of 79 possible
letter to phoneme pairings). Cluster analysis, as carried out by Rosenberg and

216 Andy Clark
Sejnowski [1987] in effect constructs a hierarchy of partitions on top of this
base level of 79 distinctive stable patterns of hidden unit activation. The
hierarchy is constructed by taking each of the 79 patterns and pairing it with
its closest neighbour, i.e. with the pattern which has most in common with it.
These pairings act as the building blocks for the next stage of analysis, in which
an average activation profile (between the members of the original pair) is
calculated and paired with its nearest neighbour drawn from the pool of
secondary figures generated by averaging each of the original pairs. The
process is repeateduntil the final pair is generated. This represents the grossest
division of the hidden unit space which the network learnt-a division, which
in the case of NETtalkturned out to correspondto the division between vowels
and consonants (see Figure 2).
Cluster analysis thus provides a kind of picture of the shape of the space of
the possible hidden unit activations which power the network's performance.
By reflecting on the various aspects of this space (i.e. the various clusterings)
the theorist can hope to obtain some insight into what the system is doing. It
may, for example, turn out to be highly sensitive to some sub-regularity which
had hitherto been unnoticed or considered unimportant. It is as if we are
providedwith a tracing of the shape of the cognitive space we are attempting to
understand. The tracing must be interpreted and that is a real and at times
difficulttask. But it is not shooting in the dark, for we can see what inputs are
associated with what configuration (even if it is a higher level configuration
revealed by cluster analysis).
We are thus given members of each class in question-the task is then to
find perspicuous, conceptual level terms in which to describethe conditions of
class membership.
A fully interpreted cluster analysis, I would like to suggest, constitutes the
nearest connectionist analogue to a classical competence theory. Like a
competence theory, it provides a level of understanding which is higher than
(i.e. more general than) the algorithmic level. For the 'algorithmic' specifica-
tion, for a connectionist, must be a specification of (a) the network
configuration and (b) the unit rules and the connection strengths. But there is
a many-one mapping between such algorithmic specifications and a particu-
lar cluster analysis. For example, a network which started out with a different
random set of weights would, after training, exhibit the same partitioning
profile (hence have an identical cluster analysis) but do so using a very
different set of individual weights. Unlike a classical competence theory,
however, the cluster analysis will typically not look like a set of state transition
rules definedover conceptual level entities. Instead, it will be more like a kind of
geometric picture of the shape of a piece of cognitive terrain. Those theorists
who think that a high level explanation must be like a set of sentences and
rules may find this hard to adjust to.
On the other hand some radically anti-sentential theorists (e.g. Churchland

Hierarchy of Partitions-4
on Hidden-Unit r t t-t
z-zt ss
d-d
Vector Space k-k
c-k
c-s
g-J
9-
j-
s-Z sZr-r
1p_ o--
aa--
ku--
w-- Y--
1- s k--
h--
h-h
W -W n-G
n-n
R
Consonants I-L
1-
c- S
l- c-C
s-S t-T
• 1 ,I ?t-D
!
t-S
-t-C
V-V
f-
m-m
p-f- P-P -
! 0 b-b
q- "'
-u-y
u-I
u-yo
- o-C
o-u
O-_
y-Y
y-i
i-Y
i-x
i-A
11e-Y
e-i
e-I
e-E
e-e •
Vowels
I e-x
e-x a-o
.....a-c
a-a a-c
a-x
FIGURE2 The results of a cluster analysis of NETtalk (from Churchland [forth-

coming-1989], after Rosenberg and Sejnowski [1987]).

218 Andy Clark
[forthcoming--1989]) may consider that an interpretedcluster analysis gives

away too much to ordinary propositional discourse. Churchland argues that
the correct level of understanding lies at the level of the connection weights.
For, he insists, those are all the system 'really' knows about; it has no
representation of its own partitionings. Moreover, the way two systems learn
given new inputs can vary even if they have identical cluster analyses at time
tl. For the connection weights (which, we saw, stand in many-one relations to
cluster analyses) are the pivotal unit of cognitive evolution.
This, however, looks like the ordinary swings and roundabouts of high level
explanation. In opting at times for a level of analysis which groups particular
connection weight specifications into equivalence classes governed by com-
mon cluster analyses we naturally trade specificityfor generality. Just as pure
Darwinism leaves recessive characteristics unexplained, but highlights
general principles covering a class of evolutionary mechanisms, so cluster
analysis leaves some details of cognitive evolution unexplained but highlights
the gross sensitivity which enables a class of networks to negotiate successfully
a given cognitive terrain. Some such high level understanding seems essential
if connectionism is to be deeply explanatory of cognitive performance. A mere
specificationof a set of connection weights is surely not an explanation,even for
the anti-sententialist.
The main point I want to stress is, however, independent of any view about
the merits or demerits of cluster analysis. It concerns the methodological
inversion of traditional cognitive science. The connectionist, by whatever
means, achieves her high level understanding of a cognitive task by reflecting
on, and tinkering with, a network which has learnt to perform the task in
question. Unlike the classical, Marr-inspiredtheorist, she does not begin with a
well worked out (sentential, symbolic) competence theory and then give it
algorithmic flesh. Instead, she begins at level -5, trains a network, and then
seeks to grasp the high level principlesit has come to embody. This is an almost
miraculous boon for cognitive science. For the discipline has been dogged by
the (related) evils of ad-hocery and sententialism. Forced to formulate
competence theories as sets of rules defined over classical, symbolic data
structures, theorists have plucked principles out of thin air to help organize
their work. Connectionist methodology, by contrast, allows the task demands
to trace themselves and thus suggest the shape of the space in a way
uncontaminated by the demands of standard symbolic formulation. We thus
avoid imposing the form of our conscious, sentential thought on our models of
unconscious processing-an imposition which was generally as practically
unsuccessful as it was evolutionarily bizarre.
In sum, the connectionist, in being compelled to make do without the
comfort of a classical competence theory is deprived neither of high-level
explanatory power nor of methodological soundness. On the contrary, the
methodology of connectionist explanation is perfectly geared to the avoidance

Connectionism, andExplanation
Competence, 219
of ad-hoc organizing principles and sentential, linguistic bias. There remain

important and unresolved questions concerning the best ways to extract and
couch such high level explanations as connectionism may provide. But
techniques such as cluster analysis, network pathology and activation
recording are already being developed and will no doubt become well-
understood. Once they do, the Copernicanrevolution in cognitive explanation
will be well under way.
7 CONCLUSIONS: THE CASCADE, THE DAM AND THE DIVIDED

STREAM
Classicists and Connectionists, it seems, must differfundamentally in the way

they expect actual processing (level 2) models to relate to traditional
competence theories. The paper began by displaying the classicist vision of this
relation and two connectionist alternatives. These may conveniently be
pictured as follows.
Relationone: the cascade

Dennett describes the classicists' vision as one of a 'triumphant cascade
through Marr'sthree levels' (Dennett [198 7], p. 22 7). The cascade flows easily
given the presence of a classical symbol processing architecture. The axioms or
rules of the competence theory are linguistically expressed formulas for
deriving one symbol from another. Various algorithms (level 2) may then
implement that derivational structure. They may do so explicitly (by tokening
the rule) or tacitly (by processing explicit symbol strings in accordance with
the rule). In this classical vision, level 2 is a neat echo of level 1.
Relationtwo: the dam

Newtonian connectionism dams the classical cascade by introducing a
dimension shift between the items (symbol strings) operated on by the level 1
derivational rules and the items (subsymbols) 'operated on' by a connectionist
network. The level 1 theory may describe some (idealized) aspects of the
network's behaviour. But the network embodies neither explicit nor tacit
knowledge of the derivational rules nor the conceptual level structures over
which they are defined.
Relationthree:the dividedstream
Rogue models represent a more complex state of affairs in which actual
performance is dependent on two systems. One, the daily, on-line system,
relates to the competence theory in the way described by the Newtonian

220 Andy Clark
connectionist, i.e. it matches some of the implied behaviour, but without
embodying the classical knowledge. The other is an additional resource,
perhaps created by the exploitation of external symbols, which simulated a
classical machine. As such, it is capable of embodying the derivational rules
and conceptual level structures specified in the competence theory. Rogue
models complicate the debate over the 'correct' architecture of cognition by
suggesting a multiplicity of interactive (virtual) architectures.
One way or another, then, the connectionist must distance herself from the
details of the classical competence model. Such models are not properly
suggestive of the form of on-line connectionist processing, though they may be
either descriptiveof (a subset of) the results of such processing, or descriptiveof
some other cognitive resource. But this dislocation of connectionism and
competence theorizing raised a serious problem. For the classicist had a
methodology which guaranteed a useful and accurate higher level under-
standing of the cognitive phenomenon modelled. The connectionist, by
contrast, may seem to have working systems but no higher level understand-
ing of them-hence, in a certain sense, no explanationsof cognitive pheno-
mena.
This worry loses some of its force once we manage to perform a kind of
Copernicanrevolution in our thinking about explanation in cognitive science.
UnderMarr'sinfluence, Cognitive Scientists are likely to expect some high level
understanding of a task to precede and inform the writing of algorithms.
Classical competence theoretic specifications aim to do just that job. The
connectionist, however, effectively inverts this strategy. She begins with a
minimal understanding of the task, trains a network to perform it, and then
seeks, in various principled ways, to achieve a higher-level understanding of
what it is doing and why. This may involve careful recording of network
activity, the examination of the network's behaviour after various forms of
damage, and plotting the way the network's hidden units divide up the
cognitive space they are negotiating. This last activity (cluster analysis, as
described in the text) clearly provides a kind of higher level understanding
since there is a many-one relation between a given cluster analysis and the set
of connection weights which could implement it. The connectionist starts with
a level *5 model, moves rapidly to a level 3 implementation and must then
work backwards to detailed higher levels of understanding.
This explanatory inversion, I want to suggest, actually constitutes one of the
major advantagesof the connectionist approach over traditional cognitive
science. It is an advantage because it providesa means by which to avoid the ad
hocgeneration of axioms and principles.Instead of having to decide on a rather
arbitraryset of symbolic, language-based axioms to organize some cognitive
task (recall naive physics) the connectionist can let the task itself organize the
network, and only thenattempt to formulate various higher level pictures of its
activity. Such pictures, moreover, may depart (in ways we have yet to fully

Connectionism, Competence, and Explanation 221
imagine) from the traditional picture of a theory as a set of propositions.

Instead, they may be more geometric, or pictorial, or may use language in
unexpected, apparently clumsy ways (see Churchland [forthcoming-198 9]).
There is, we may finally conjecture, a fairly deep reason why attitudes to
competence polarize connectionists and classicists. It is that a competence
model is a traditional theory, expressed in propositional or logical form.
Classicists believe that thinking just is the manipulation of items having
propositional or logical form;connectionists insist that this is just the icing on
the cake and that thinking ('deep' thinking, rather than just sentence
rehearsal) depends on the manipulation of quite differentkinds of structure. As
a result, the classicist attempts to give a level 2 processing model which is
defined over the very same kinds of structure as figure in her level 1 theory.
Whereas the connectionist insists on dissolving that structure and replacing it
with something quite different.
A curious irony emerges. In the early days of Artificial Intelligence, the
rallying cry was 'Computersdo not crunch numbers, they manipulatesymbols!'
This was meant to inspire a doubting public by showing how much
computation was like thinking. Now the wheel has come full circle. The virtue
of connectionist systems, it seems, is that 'they do not manipulate symbols,
they crunchnumbers'.And nowadays we all know (don't we?) that thinking is
not mere symbol manipulation! So the wheel turns.
School of Cognitive & Computing Sciences
University of Sussex
Brighton
REFERENCES
CHOMSKY, N. [1986]: Knowledgeof Language:Its Nature, Origin and Use, Praeger
Publishers, Connecticut.
CHURCHLAND, P. [forthcoming--1989]: 'On the nature of theories: a neurocomputa-
tional perspective'. In P. M. Churchland (ed.) TheNeurocomputational Perspective,
MIT Press, Cambridge,Massachusetts.
CLARK,A. [1988]: 'Thoughts, sentences and cognitive science', PhilosophicalPsycho-
logy, Vol. I, no. 3, pp. 263-78.
CLARK,A. [1989]: Microcognition:Philosophy, CognitiveScienceand ParallelDistributed
Processing.MIT/BradfordBooks, Cambridge, Massachusetts.
DAVIES,M. [198 7]: 'Tacit knowledge and semantic theory: can a five per cent difference
matter?', Mind, 96, pp. 441-62.
DAVIES,M. [forthcoming]: 'Connectionism, modularity and tacit knowledge', British
Journalfor the Philosophyof Science.
DENNETrD, [198 7]: TheIntentionalStance.MIT/BradfordBooks, Cambridge,Massachu-
setts.
DENNETr,D. [1988]: 'The evolution of consciousness', JacobsenLecture,University of
London, May 1988. Tufts UniversityCurrentCirculatingManuscriptCCM-88-1.

222 Andy Clark
DENNETT, D. [forthcoming]: 'Review of Psychosemantics', Journalof Philosophy.
FODOR, J. and PYLYSHYN, Z. [1988]: 'Connectionism and cognitive architecture',
Cognition,28, pp. 3-71.
HAYES, P. [1984]: 'Liquids'in Formal Theoriesof the CommonsenseWorlded. J. Hobbs
(Ablex, Hillsdale, NJ, 1984).
KARMILOFF-SMITH, A. [1987]: 'Beyond modularity: a developmental perspective on
human consciousness'. Transcript of talk given to the Annual Meeting of the
British Psychological Society, Sussex, April 1987.
MARR,D. [1977]: 'ArtificialIntelligence: a personal view', In J. Haugeland (ed.) Mind
Design. MIT/BradfordBooks, Cambridge,Massachusetts, 1981.
PEACOCKE, C. [1986]: 'Explanationin computational psychology: language, perception
and level Mind and Language,Vol. 1, No. 2, pp. 101-23.
PINKER, A. and1.5',
PRINCE, S. [1988]: 'On language and connectionism', Cognition28.
RIDLEY, M. [1985]: TheProblemsof Evolution.OxfordUniversity Press, Oxford.
ROSENBERG, C. and SEJNOWSKI, T. [1987]: 'Parallel networks that learn to pronounce
English text', ComplexSystems, I, pp. 145-68.
RUMELHART, D. and MCCLELLAND, J. [1986]: 'PDPmodelsandgeneralissuesin cognitive
science'. In J. McClelland, D. Rumelhart and the PDP Research Group (eds.),
ParallelDistributedProcessing:Explorationsin the Microstructureof Cognition.MIT/
BradfordBooks, Cambridge,Massachusetts, 1986, Vol. I, pp. 110-46.
RUMELHART, D. and MCCLELLAND, J. [1986b): 'On learning the past tenses of English
verbs', In J. McClelland,D. Rumelhart and the PDP Research Group (eds.), Parallel
DistributedProcessing:Explorationsin the Microstructureof Cognition.MIT/Bradford
Books, Cambridge,Massachusetts, 1986, Vol. II, pp. 216-71.
RUMELHART, D., SMOLENSKY,P., MCCLELLAND,J., and HINTON, G. [1986]: Schemata and
sequential thought processes in PDP models. In J. McClelland,D. Rumelhart and
the PDP Research Group (eds.), ParallelDistributedProcessing:Explorationsin the
Microstructureof Cognition.MIT/BradfordBooks, 1986, Cambridge, Massachu-
setts, Vol. II, pp. 7-57.
SEJNOWSKI, T. and ROSENBERG,C. [1986]. 'NETtalk:a parallel network that learns to
read aloud'. Johns Hopkins UniversityElectricalEngineeringand ComputerScience
TechnicalReport.JHU/EEC-86/01.
SMOLENSKY,P. [1986]: Information processing in dynamical systems: foundations of
harmony theory. In J. McClelland, D. Rumelhart and the PDP Research Group
(eds.), ParallelDistributedProcessing:Explorationsin the Microstructureof Cognition.
MIT/BradfordBooks 1986, Cambridge,Massachusetts, Vol. I, pp. 194-281.
SMOLENSKY, P. [1988]: 'Onthe propertreatment of connectionism', BehavioralandBrain
Sciences,11, pp. 1-73.
SMOLENSKY, P. [1987]: 'The constituent structure of connectionist mental states',
SouthernJournalof Philosophy,Vol. XXVI,Supp. pp. 137-62.


10 - Clark (Eng) PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

10 - Clark (Eng) PDF

Uploaded by

Copyright:

Available Formats

The British Society for the Philosophy of Science

Connectionism, Competence, and Explanation

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

Connectionism, Competence, and

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

2 LEVELS OF EXPLANATION AND THE IDEA OF AN EQUIVALENCE

Explanation,it seems,is a many-levelledthing.A singlephenomenonmay be

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

FIGURE 1 The swings and roundaboutsof explanatorygenerality.WhatDNA-based

the why of a computation, this specification can be progressively refined so as

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

3 THE CLASSICAL CASCADE

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

implementations. Indeed, it begins to seem as if that close relation is what is

If (A and B) then (A)

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

The connectionist vision of the relation between a structured competence

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

To illustrate this point, consider a simple example due to Paul Smolensky.

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

6 THE METHODOLOGY OF CONNECTIONIST EXPLANATION

Connectionist explanatory strategies, it seems, cannot fit into the mould

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

instances, to negotiate its way around this particularly tricky cognitive

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

FIGURE2 The results of a cluster analysis of NETtalk (from Churchland [forth-

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

[forthcoming--1989]) may consider that an interpretedcluster analysis gives

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

of ad-hoc organizing principles and sentential, linguistic bias. There remain

7 CONCLUSIONS: THE CASCADE, THE DAM AND THE DIVIDED

Classicists and Connectionists, it seems, must differfundamentally in the way

Relationone: the cascade

Relationtwo: the dam

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

imagine) from the traditional picture of a theory as a set of propositions.

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

This content downloaded from 201.234.181.53 on Thu, 5 Feb 2015 22:08:48 PM

You might also like